One of humanity's top unsolved tech challenges has been the ability to read each other's minds, know what others think.
It just dawned on me that we have a solution for a big part of that today.
With DeepSeek-R1, the reasoning model, that became freely available last week. $^1$
Think about this for a second.
We have so far mever been able to know the exact thought process of any other intelligent being.
We definitely have no idea what goes on in any other person's mind.
Even for ourselves, many things are so spontaneous, or complex, we don't know exactly how we think.
For the first time now, we can see the thought process of a very intelligent AI that is trained to think well and share its thought process.
Recent models have also been able to think.
The early thought processes were rudimentary and crude. And then, there were good ones like o1 and the upcoming o3, but they chose not to share the full thought process with users. There were others, but they were not widely available from a browser and could only be run by programmers on powerful machines - qwq and marco-o1 are two examples.
DeepSeek-R1 is the first widely available model that openly shared its thought process, and in many cases, it is extremely useful and relatable to how we might think about the same problem.
But, these models are from China
If you are in the US or in some other places, the fact that DeepSeek and Qwen chat are from China may be a concern. I personally don't care where the models are from for anything I have to chat with AI for:
I don't care which country knows the kind of things I want to talk about with AI.
If I really had something personal I don't want an AI system to know, I would use some AI with temporary chat to reduce the spread of such info. Or use local models on my own machine - see my local models chat. $^3$
Of course, if you have national security or business related reasons to avoid your data going to China, then use other chat models.
And if you have similar concerns with accessing these programmatically, then either use them with a host that hosts them locally in the US that meets your needs, or on a secure local machine (see here for more $^3$).
Appendix - My current AI chat from browser stack
For any task, try each in order, skip if one does not support what you want, or does not give a good response:
Deepseek-R1 > Qwen2.5-Max $^2$ > claude 3.5 sonnet > o1 > gpt-4o > gemini 1.5 pro with deep research
Today, I added Qwen Chat to this list, as it does everything I need from ChatGPT or Claude.
Qwen is behind Deepseek R1, which I added 3 days ago, only because I love to see Deepseek R1 think. Also, the chat interface of DeepSeek chat gives me R1 combined with web search, which ChatGPT o1 did not provide me. Qwen's best model Qwen2.5-Max, does not share its reasoning, but does pretty well on many queries. Qwen chat also has reasoning models like QWQ that share their reasoning, and the reasoning is pretty good, but I need to compare it more with Deepseek R1's reasoning to decide.
Claude is still a very intelligent model, better than ChatGPT for many complex tasks (including coding) but it cannot search the web from within the browser UI. And it also does not share its thought process, a new feature that is becoming very useful.
Support me: The best way to support my continued work (articles here, research, new AI based experiences) is via ko-fi. Please let me know your feedback, and what you would like to see more of in the future. Thank you!
$1.$ Try DeepSeek-R1 and other DeepSeek models here for free, be sure to select the DeepSeek-R1 by clicking on it. When we ask this model something, it first explains, within <think></think> tags, its entire thought process. Then it proceeds to answer the question.
$2.$ Try Qwen2.5 models here for free, be sure to select the Qwen2.5-Max model for the best response and the widest set of capabilities, or one of the reasoning models like QwQ if you want to see its thought process, with possibly lower quality responses and fewer capabilities.
$3.$ See my comparison of 98 local models on a custom dataset here